Binding Pose Prediction Eval

binding-pose-prediction-eval

Geometric Deep Learning for Structure-Based Drug Design: A Survey — Zaixi Zhang et al. (2023) (arXiv:2306.11768, 2023)

What this evaluates

Evaluates a model's ability to predict the 3D spatial arrangement (binding pose) of a small molecule ligand when docked to a target protein structure. It probes geometric reasoning, conformational sampling, and the capacity to generate physically plausible protein-ligand complexes.

Datasets

• PDBBind — total 17347; splits: train (-1), test (-1)

Metrics

• L-RMSD (primary) — range: other
  • Ligand Root Mean Square Deviation (L-RMSD) measures the mean squared error between the atoms of the predicted and ground truth ligand coordinates. Computed as sqrt((1/n) * sum(||R_i - R_hat_i||^2)) for i=1 to n, where n is the number of atoms.
• Centroid Distance — range: other
  • Calculates the Euclidean distance between the averaged coordinates of the predicted ligand atoms and the truly bound ligand atoms.
• Kabsch RMSD — range: other
  • Computes the lowest possible RMSD by first applying the Kabsch algorithm to optimally superimpose the predicted and ground truth structures via roto-translation, then calculating the standard RMSD.

Input / output format

Input: 3D structural representation of the target protein (receptor) and the 2D graph or 3D structure of the ligand.

Output: 3D coordinates of the ligand atoms representing the predicted binding pose relative to the protein.

Scoring recipe

def compute_l_rmsd(predicted_coords, true_coords):
    # predicted_coords and true_coords are (n, 3) numpy arrays
    n = predicted_coords.shape[0]
    diff = predicted_coords - true_coords
    sq_dist = np.sum(diff**2, axis=1)
    return np.sqrt(np.mean(sq_dist))

def compute_kabsch_rmsd(predicted_coords, true_coords):
    # Align predicted_coords to true_coords using Kabsch algorithm
    aligned = kabsch_align(predicted_coords, true_coords)
    return compute_l_rmsd(aligned, true_coords)

Common pitfalls

• Treating proteins as rigid bodies during docking, ignoring inherent conformational flexibility that occurs upon ligand binding.
• Focusing exclusively on geometric accuracy (e.g., low RMSD) while neglecting chemical and physical plausibility, leading to poses with steric clashes or invalid bond geometries.

Evidence (verbatim from paper)

Ligand Root Mean Square Deviation (L-RMSD) is the mean squared error between the atoms of the predicted and bound ligands. Formally, let $R\in\mathbb{R}^{n\times 3}$ and $\hat{R}\in\mathbb{R}^{n\times 3}$ be the predicted and the ground truth ligand coordinates, where $n$ is the number of atoms. The L-RMSD is obtained with: $\text{L-RMSD}(R,\hat{R})=\big{(}\frac{1}{n}\sum_{i=1}^{n}||R_{i}-\hat{R}_{i}||^{2}\big{)}^{\frac{1}{2}}$

Citation

@misc{zhang2023geometric,
  title={Geometric Deep Learning for Structure-Based Drug Design: A Survey},
  author={Zaixi Zhang et al. (2023)},
  year={2023},
  note={arXiv:2306.11768}
}

• arXiv: 2306.11768